Multi-Modal Pancreas And Pancreatic Tumor Image Segmentation

Pancreatic cancer is the only disease in which the number of morbidity and death is close to 1:1,which seriously endangers human life and health.Nowadays,computer-aided diagnosis technology has received more and more attention.Accurate segmentation of pancreas and tumor is very important for the detection and radiotherapy of computer-aided diagnosis of pancreatic cancer.However,because the pancreas and its tumor lesions are highly similar to the surrounding tissues,and their size,shape and location are variable,achieving accurate segmentation is still one of the most challenging tasks in the field of medical image segmentation.The focus of this article is on the automatic segmentation of pancreas and tumors in Magnetic Resonance,which is very important for pancreatic MRI-guided radiotherapy.However,MRI images alone may not be able to accurately define the boundaries of pancreatic tumors.On the other hand,Positron Emission Tomography images can obtain high contrast between tumors and normal tissues,thereby reducing differences in tumor-localization between patients and providing further help for the diagnosis and treatment of the pancreas.Therefore,it is necessary to study the segmentation of pancreas and tumor in multi-modal images.In this paper,based on PET/MRI pancreatic image data,using the current popular deep learning methods,we propose a multi-modal pancreas image segmentation algorithm.The main research content of this article includes the following three aspects:1.Aiming at the problem of blurred MRI boundaries and the difficulty of accurate positioning and segmentation due to large differences in pancreatic organs,a MRI pancreatic image segmentation method based on shape constraints and multi-scale networks is proposed.First,the pre-trained U-net model is used to learn the position information of pancreatic tumors through PET images,and the shape expression model constructed by the stacked convolutional autoencoder is used to learn the shape information of the pancreas.Then a multi-scale network model for MRI pancreatic image segmentation was designed.The pre-trained U-net model and shape expression model were used to further adjust the results by constraining the segmentation model parameter optimization direction.Finally,the final segmentation result can be directly obtained by inputting the MRI image into the trained multi-scale model.The algorithm is evaluated through experiments,and the effectiveness of the shape expression model and multi-scale modules in improving the accuracy of network segmentation is verified.2.Aiming at the problem of indefinite size and location of pancreatic tumor lesions and poor segmentation results for small targets,a multi-modal pancreatic tumor segmentation method based on attention mechanism and level set is proposed.First,build a multi-scale network based on the attention mechanism.The network uses two paths to extract the global and local features of different images.At the same time,it uses the attention mechanism and the multi-scale convolution module to learn complementary information in different fields and suppress irrelevant information,making the network more sensitive to the edge of the tumor.Input the registered PET and MRI images into the network to obtain the position and approximate outline of the pancreatic tumor.The rough segmentation results are post-processed using morphological methods,and the final segmentation results are automatically evolved through the level set model of distance regularization.Experimental results show that this segmentation method can accurately locate pancreatic tumors and improve the segmentation effect to a certain extent.3.In view of the low contrast of MRI images and the problem of under-segmentation and false positive regions in tumor segmentation results,a pancreatic tumor detection and segmentation method based on Dense module and migration learning is proposed.The feature extraction network is improved based on the Mask-RCNN network structure.The Dense module and the feature pyramid structure are combined,and the extracted multi-scale features are reused to detect the tumor area more accurately.Through migration learning,the underlying network can share features from the tumor in the PET image,which alleviates the problem of insufficient training samples,reduces network overfitting,and achieves accurate segmentation of pancreatic tumors.Finally,through comparative experiments,the necessity and effectiveness of each module are verified.Our method has excellent performance on visual effects and indicators.